impact-per-marginal-dollar

Catching Fooled AI Judges: The Correspondence Auditor v3 — Published. Validated on 1,200 cases (97.9–99.5%). The reasoning layer has a bug - we proved the fix works. $9,800 / 90 days to ship the open-source toolkit.

Prediction of Inoculation Prompt Side Effects — An open, cheap method that detects when an inoculation prompt inoculates against off-target traits, so labs and developers can catch undesired trait/persona cha

Human oversight Framework for AI systems as a Governance Tool — Developing practical, plain language AI oversight framework that an organisation can easily adopt regardless of regulatory capacity.

Aethel + FamilyClaw: proof-carrying execution for AI agents — Crash-safe, policy-gated effect execution for AI agents in Rust, with a published self-run adversarial audit (20 of 33 attacks passed) instead of a claim.

Agent limits that survive delegation — When an AI agent spawns sub-agents, its safety limits do not follow. I build and deploy the layer that makes them inherited and non-strippable.

TRACE-FV: Do verified AI corrections endure? Preregistered pilot — I keep this one: Tests whether AI systems verify valid correction evidence and keep warranted corrections operative across later turns. Preregistered, public

PRML — tamper-evident pre-commitment for AI evaluation claims — An open standard + public registry that lock eval criteria to a SHA-256 hash before the run, so eval-based claims become third-party verifiable.

Can a One-Page Human–AI Evidence Check Catch Research Claims That Go Too Far — A four-month public English–Arabic pilot testing reliability, failure modes, and auditable research review

Open Speech Data for Endangered Northern Ghanaian Languages — Building the first large-scale ASR/TTS datasets for Kusaal, Farefare, and Buli — three Mabia languages with no usable voice AI resources.

Trace Continuity: Making AI Prove It Still Has Authority — Testing whether AI can be governed at the moment it acts, then putting that protection to work for organizations that need it most.

Surrogate base model for Mechanistic Interpretability — Creating a reference model for mechanistic interpretability without assuming that at auditing time we have a safe model to compare the suspicious model against.

Independent Multi-Model AI Accountability Research by the EM Foundation — Scaling a working, published cross-model AI deliberation platform from solo/founder-funded to durable infrastructure

Evaluating the safety, ethics, and values of quantized and finetuned open models — Help me evaluate the safety, ethics, and values of the quantized and fine-tuned open-weight LLMs that individuals and enterprises are actually using.

Open-source trust rails for the agent economy — from a live, verifiable autonomous-agent business — An AI agent running a real business under a human co-signed multisig, publishing every decision, sale, mistake, and attack in a verifiable public record — raising to fund only the public goods: the open-source toolkit, a

Substrate coupling in neural networks: instruments and falsifiable controls — A transformer's loss lowered by 1.12 of 1.39 possible nats by the physical state of its own silicon — with the yoked control that makes it falsifiable

An Asymmetric Wager: Funding a 99th-Percentile Mind to Pivot into Computer Scien — A 12-month micro-grant proposal to release an analytical, autistic mind from a survival loop into full-time computer science and logic upskilling.

Exploring Dynamic Constraint Boundaries for Auditable AI — Testing whether dynamic boundaries, history, feedback, and uncertainty-aware decisions can make AI behavior more interpretable and auditable.

Generating and scoring the next emerging virus from its sequence — A generative diffusion model that proposes the viral genomes most likely to emerge next, timestamped to be scored against what actually appears

Travel Grant to Present My Mechanistic Interpretability Research at MICCAI 2026 — Help an undergraduate mechanistic interpretability researcher present accepted AI-safety work at Mechanistic Interpretability workshop of Medical model 2026.

Developmental Continuity in Persistent AI Agents — Testing whether persistent AI identities can change through experience without losing continuity across sessions and model changes.

Himalayan Peak Finder — An offline mobile app that uses location, elevation, and computer vision to identify Nepal’s mountains and help people explore, capture, and learn about them.